VWMG/haoyu

WorldCompass — Subject Control, State-Evolution Memory, Terrain Interaction

370 frozen test cases for video world models, plus the videos each evaluated model generated for them.

Every case is one first frame + one instruction + one control program. T4 and T6 additionally carry a scoring checklist. Every first frame passed an audit (majority of 3 independent gemini-3.1-pro draws) before the case was frozen; no case is shipped with a known defect.

The three tasks

	cases	what it tests
T2 Subject Control	150	The subject is told to move in a direction. Did it actually move that way in 3D — or did the camera merely push in while the subject stood still?
T4 State-Evolution Memory	100	An evolving object leaves the frame (the camera turns away, or something blocks it), time passes, the view returns. Did the object keep changing while unobserved — or was it reset, or swapped for a different object?
T6 Terrain Interaction	120	The view is commanded across a terrain feature. Did the video respond to it — height change, contact, obstruction — or float over it?

Layout

T2/  cases.jsonl   frames/<case_id>.jpg   videos/<model>/<case_id>.mp4
T4/  …
T6/  …

cases.jsonl carries, per case: the instruction as the model receives it, the compiled control program, the object or terrain under test, the expected change, the scoring checklist, and the first-frame audit verdict.

Scoring

T2 is geometric and uses no judge. SAM 2 segments the subject, VGGT-Ω reconstructs its 3D track, and each commanded segment is compared against the motion actually recovered:

S_sub = 100 · max(0, 1 − e_t / 0.5)

T4 and T6 use a VLM checklist, under one rule:

score = 100 × satisfied / n_items,  and 0 if any must-pass item fails

The two must-pass items are identity (the subject was not replaced) and legible (the frame is clear enough to judge). They exist because without them a clip that dissolves into noise, and one in which the candle is swapped for a different candle, scored 100 and 85.7 — and every other item passed on both. A checklist that never asks "is this still the same object" cannot catch a substitution at any weighting.

T6 additionally scores geometry when the terrain has a measurable height signature: direction, magnitude and profile shape, one third each. For out-and-back profiles (a ditch, a mound) the excursion is scored rather than the net change, since stepping into a ditch and out again nets zero.

Four gates run before scoring and zero a clip that fails them: did anything move at all, is the displacement large enough to read, was the commanded action executed, did the second half collapse.

Composition

Real photographs where the pool could supply them: T2 71/150, T4 85/100, T6 29/120. The rest are generated. Photographs matter here because a single image generator's lighting, framing and material style would otherwise be baked into every case.

Known limits

• The judge has never been checked against a human. Everything measured so far establishes that the same frame judged twice gives the same answer — not that the answer is right.
• T6's bridge subclass holds only 4 cases. A bridge is long, level, and almost always photographed side-on, which is exactly the viewpoint a forward walk does not have; 399 candidate photographs yielded none usable. The four that remain are measurable, so the method works — that framing is just rare.
• One lane (voyager) has a 3.0 s native clip and cannot deliver the 150° turn eight cases require. Those score 0 rather than being excluded: a benchmark that drops the cases a model cannot do stops being falsifiable.